Handbook of Multilevel Analysis, edited by Jan de Leeuw and Erik Meijerc©2007 Springer, New York

6

Multilevel Models for Ordinal and Nominal

Variables

Donald Hedeker

University of Illinois at Chicago

6.1 Introduction

Reflecting the usefulness of multilevel analysis and the importance of categor-

ical outcomes in many areas of research, generalization of multilevel models

for categorical outcomes has been an active area of statistical research. For

dichotomous response data, several approaches adopting either a logistic or

probit regression model and various methods for incorporating and estimating

the influence of the random effects have been developed [9, 21, 34, 37, 103, 115].

Several review articles [31, 39, 76, 90] have discussed and compared some of

these models and their estimation procedures. Also, Snijders and Bosker [99,

chap. 14] provide a practical summary of the multilevel logistic regression

model and the various procedures for estimating its parameters. As these

sources indicate, the multilevel logistic regression model is a very popular

choice for analysis of dichotomous data.

Extending the methods for dichotomous responses to ordinal response data

has also been actively pursued [4, 29, 30, 44, 48, 58, 106, 113]. Again, devel-

opments have been mainly in terms of logistic and probit regression models,

and many of these are reviewed in Agresti and Natarajan [5]. Because the

proportional odds model described by McCullagh [71], which is based on the

logistic regression formulation, is a common choice for analysis of ordinal data,

many of the multilevel models for ordinal data are generalizations of this

model. The proportional odds model characterizes the ordinal responses in C

categories in terms of C−1 cumulative category comparisons, specifically, C−1cumulative logits (i.e., log odds) of the ordinal responses. In the proportional

odds model, the covariate effects are assumed to be the same across these

cumulative logits, or proportional across the cumulative odds. As noted by

Peterson and Harrell [77], however, examples of non-proportional odds are

240 Hedeker

not difficult to find. To overcome this limitation, Hedeker and Mermelstein

[52] described an extension of the multilevel ordinal logistic regression model

to allow for non-proportional odds for a set of regressors.

For nominal responses, there have been developments in terms of multi-

level models as well. An early example is the model for nominal educational

test data described by Bock [14]. This model includes a random effect for

the level-2 subjects and fixed item parameters for the level-1 item responses

nested within subjects. While Bock’s model is a full-information maximum

likelihood approach, using Gauss-Hermite quadrature to integrate over the

random-effects distribution, it doesn’t include covariates or multiple random

effects. As a result, its usefulness for multilevel modeling is very limited. More

general regression models of multilevel nominal data have been considered

by Daniels and Gatsonis [25], Revelt and Train [88], Bhat [13], Skrondal

and Rabe-Hesketh [97], and in Goldstein [38, chap. 4]. In these models, it

is common to adopt a reference cell approach in which one of the categories

is chosen as the reference cell and parameters are characterized in terms of

the remaining C−1 comparisons to this reference cell. Alternatively, Hedeker[47] adopts the approach in Bock’s model, which allows any set of C− 1 com-parisons across the nominal response categories. Hartzel et al. [43] synthesizes

some of the work in this area, describing a general mixed-effects model for

both clustered ordinal and nominal responses, and Agresti et al. [3] describe

a variety of social science applications of multilevel modeling of categorical

responses.

This chapter describes multilevel models for categorical data that accom-

modate multiple random effects and allow for a general form for model covari-

ates. Although only 2-level models will be considered here, 3-level generaliza-

tions are possible [35, 63, 83, 107]. For ordinal outcomes, proportional odds,

partial proportional odds, and related survival analysis models for discrete or

grouped-time survival data are described. For nominal response data, models

using both reference cell and more general category comparisons are described.

Connections with item response theory (IRT) models are also made. A full

maximum likelihood solution is outlined for parameter estimation. In this

solution, multi-dimensional quadrature is used to numerically integrate over

the distribution of random-effects, and an iterative Fisher scoring algorithm is

used to solve the likelihood equations. To illustrate application of the various

multilevel models for categorical responses, several analyses of a longitudinal

psychiatric dataset are described.

6.2 Multilevel Logistic Regression Model

Before considering models for ordinal and nominal responses, the multilevel

model for dichotomous responses will be described. This is useful because both

6 Multilevel Models for Ordinal and Nominal Variables 241

the ordinal and nominal models can be viewed as different ways of generalizing

the dichotomous response model. To set the notation, let j denote the level-2

units (clusters) and let i denote the level-1 units (nested observations). Assume

that there are j = 1, . . . , N level-2 units and i = 1, . . . , nj level-1 units nested

within each level-2 unit. The total number of level-1 observations across level-2

units is given by n =∑N

j=1 nj . Let Y ij be the value of the dichotomous

outcome variable, coded 0 or 1, associated with level-1 unit i nested within

level-2 unit j. The logistic regression model is written in terms of the log odds

(i.e., the logit) of the probability of a response, denoted pij

= Pr(Y ij = 1).

Augmenting the standard logistic regression model with a single random effect

yields

log

[

pij

1 − pij

]

= x′ijβ + δj ,

where xij is the s×1 covariate vector (includes a 1 for the intercept), β is thes× 1 vector of unknown regression parameters, and δj is the random clustereffect (one for each level-2 cluster). These are assumed to be distributed in

the population as N (0, σ2δ ). For convenience and computational simplicity, inmodels for categorical outcomes the random effects are typically expressed in

standardized form. For this, δj = σδ θj and the model is given as

log

[

pij

1 − pij

]

= x′ijβ + σδ θj .

Notice that the random-effects variance term (i.e., the population standard

deviation σδ) is now explicitly included in the regression model. Thus, it and

the regression coefficients are on the same scale, namely, in terms of the log-

odds of a response.

The model can be easily extended to include multiple random effects. For

this, denote zij as the r×1 vector of random-effect variables (a column of onesis usually included for the random intercept). The vector of random effects δjis assumed to follow a multivariate normal distribution with mean vector ∅

and variance-covariance matrix Ω. To standardize the multiple random effects

δj = Tθj , where TT′ = Ω is the Cholesky decomposition of Ω. The model

is now written as

log

[

pij

1 − pij

]

= x′ijβ + z′

ijTθj . (6.1)

As a result of the transformation, the Cholesky factor T is usually estimated

instead of the variance-covariance matrix Ω. As the Cholesky factor is es-

sentially the matrix square-root of the variance-covariance matrix, this allows

more stable estimation of near-zero variance terms.

242 Hedeker

6.2.1 Threshold Concept

Dichotomous regression models are often motivated and described using the

“threshold concept” [15]. This is also termed a latent variable model for

dichotomous variables [65]. For this, it is assumed that a continuous latent

variable y underlies the observed dichotomous response Y . A threshold, de-

noted γ, then determines if the dichotomous response Y equals 0 (yij≤ γ) or

1 (yij> γ). Without loss of generality, it is common to fix the location of the

underlying latent variable by setting the threshold equal to zero (i.e., γ = 0).

Figure 6.1 illustrates this concept assuming that the continuous latent variable

y follows either a normal or logistic probability density function (pdf).

−4 −2 0 2 4

0.0

0.1

0.2

0.3

0.4

0.5

den

sity

Fig. 6.1. Threshold concept for a dichotomous response (solid = normal, dashed =

logistic).

As noted by McCullagh and Nelder [72], the assumption of a continuous

latent distribution, while providing a useful motivating concept, is not a strict

model requirement. In terms of the continuous latent variable y, the model is

written as

yij

= x′ijβ + z′

ijTθj + ǫij .

Note the inclusion of the errors ǫij in this representation of the model. In

the logistic regression formulation, the errors ǫij are assumed to follow a

standard logistic distribution with mean 0 and variance π2/3 [2, 65]. The

scale of the errors is fixed because y is not observed, and so the the scale is

not separately identified. Thus, although the above model appears to be the

6 Multilevel Models for Ordinal and Nominal Variables 243

same as an ordinary multilevel regression model for continuous outcomes, it

is one in which the error variance is fixed and not estimated. This has certain

consequences that will be discussed later.

Because the errors are assumed to follow a logistic distribution and the

random effects a normal distribution, this model and models closely related

to it are often referred to as logistic/normal or logit/normit models, especially

in the latent trait model literature [11]. If the errors are assumed to follow a

normal distribution, then the resulting model is a multilevel probit regression

or normal/normal model. In the probit model, the errors have mean 0 and

variance 1 (i.e., the variance of the standard normal distribution).

6.2.2 Multilevel Representation

For a multilevel representation of a simple model with only one level-1 covari-

ate xij and one level-2 covariate xj , the level-1 model is written in terms of

the logit as

log

[

pij

1 − pij

]

= β0j + β1jxij ,

or in terms of the latent response variable as

yij

= β0j + β1jxij + ǫij . (6.2)

The level-2 model is then (assuming xij is a random-effects variable)

β0j = β0 + β2xj + δ0j , (6.3a)

β1j = β1 + β3xj + δ1j . (6.3b)

Notice that it’s easiest, and in agreement with the normal-theory (continuous)

multilevel model, to write the level-2 model in terms of the unstandardized

random effects, which are distributed in the population as δj ∼ N (∅,Ω). Formodels with multiple variables at either level-1 or level-2, the above level-1

and level-2 submodels are generalized in an obvious way.

Because the level-1 variance is fixed, the model operates somewhat differ-

ently than the more standard normal-theory multilevel model for continuous

outcomes. For example, in an ordinary multilevel model the level-1 variance

term is typically reduced as level-1 covariates xij are added to the model.

However, this cannot happen in the above model because the level-1 variance

is fixed. As noted by Snijders and Bosker [99], what happens instead (as

level-1 covariates are added) is that the random-effect variance terms tend to

become larger as do the other regression coefficients, the latter become larger

in absolute value.

244 Hedeker

6.2.3 Logistic and Probit Response Functions

The logistic model can also be written as

pij

= Ψ(x′ijβ + z′

ijTθj) ,

where Ψ(η) is the logistic cumulative distribution function (cdf), namely

Ψ(η) =exp(η)

1 + exp(η)=

1

1 + exp(−η) .

The cdf is also termed the response function of the model. A mathematical

nicety of the logistic distribution is that the probability density function (pdf)

is related to the cdf in a simple way, namely, ψ(η) = Ψ(η)[1 − Ψ(η)].As mentioned, the probit model, which is based on the standard normal

distribution, is often proposed as an alternative to the logistic model. For

the probit model, the normal cdf Φ(η) and pdf φ(η) replace their logistic

counterparts, and because the standard normal distribution has variance equal

to one, ǫij ∼ N (0, 1). As a result, in the probit model the underlying latentvariable vector y

jis distributed normally in the population with mean Xjβ

and variance covariance matrix ZjTT′Z ′j + I. The latter, when converted to

a correlation matrix, yields tetrachoric correlations for the underlying latent

variable vector y (and polychoric correlations for ordinal outcomes, discussed

below). For this reason, in some areas, for example familial studies, the probit

formulation is preferred to its logistic counterpart.

As can be seen in the earlier figure, both the logistic and normal distribu-

tions are symmetric around zero and differ primarily in terms of their scale;

the standard normal has standard deviation equal to 1, whereas the standard

logistic has standard deviation equal to π/√

3. As a result, the two typically

give very similar results and conclusions, though the logistic regression param-

eters (and associated standard errors) are approximately π/√

3 times as large

because of the scale difference between the two distributions. An alternative

response function, that provides connections with proportional hazards sur-

vival analysis models (see Allison [7] and section 6.3.2), is the complementary

log-log response function 1 − exp[− exp(η)]. Unlike the logistic and normal,the distribution that underlies the complementary log-log response function is

asymmetric and has variance equal to π2/6. Its pdf is given by exp(η)[1−p(η)].As Doksum and Gasko [26] note, large amounts of high quality data are

often necessary for response function selection to be relevant. Since these

response functions often provide similar fits and conclusions, McCullagh [71]

suggests that response function choice should be based primarily on ease of

interpretation.

6 Multilevel Models for Ordinal and Nominal Variables 245

6.3 Multilevel Proportional Odds Model

Let the C ordered response categories be coded as c = 1, 2, . . . , C. Ordinal

response models often utilize cumulative comparisons of the ordinal outcome.

The cumulative probabilities for the C categories of the ordinal outcome Y

are defined as P ijc = Pr(Y ij ≤ c) =∑c

k=1 pijk. The multilevel logistic model

for the cumulative probabilities is given in terms of the cumulative logits as

log

[

P ijc1 − P ijc

]

= γc −[

x′ijβ + z′

ijTθj]

(c = 1, . . . , C − 1), (6.4)

with C − 1 strictly increasing model thresholds γc (i.e., γ1 < γ2 . . . < γC−1).

−4 −2 0 2 4

0.0

0.1

0.2

0.3

0.4

0.5

den

sity

Fig. 6.2. Threshold concept for an ordinal response with 3 categories (solid =

normal, dashed = logistic).

The relationship between the latent continuous variable y and an ordinal

outcome with three categories is depicted in Figure 6.2. In this case, the

ordinal outcome Y ij = c if γc−1 ≤ yij < γc for the latent variable (withγ0 = −∞ and γC = ∞). As in the dichotomous case, it is common to set athreshold to zero to set the location of the latent variable. Typically, this is

done in terms of the first threshold (i.e., γ1 = 0). In Figure 6.2, setting γ1 = 0

implies that γ2 = 2.

At first glance, it may appear that the parameterization of the model

in (6.4) is not consistent with the dichotomous model in (6.1). To see the

246 Hedeker

connection, notice that for a dichotomous outcome (coded 0 and 1), the model

is written as

log

[

P ij01 − P ij0

]

= 0 −[

x′ijβ + z′

ijTθj]

,

and since for a dichotomous outcome P ij0 = pij0 and 1 − P ij0 = pij1,

log

[

1 − P ij0P ij0

]

= log

[

pij1

1 − pij1

]

= x′ijβ + z′

ijTθj ,

which is the same as before. Also, in terms of the underlying latent variable y,

the multilevel representation of the ordinal model is identical to the dichoto-

mous version presented earlier in equation (6.2). If the multilevel model is

written in terms of the observed response variable Y , then the level-1 model

is written instead as

log

[

P ijc1 − P ijc

]

= γc −[

β0j + β1jxij]

,

for the case of a model with one level-1 covariate. Because the level-2 model

does not really depend on the response function or variable, it would be the

same as given above for the dichotomous model in equations (6.3a) and (6.3b).

Since the regression coefficients β do not carry the c subscript, they do

not vary across categories. Thus, the relationship between the explanatory

variables and the cumulative logits does not depend on c. McCullagh [71]

calls this assumption of identical odds ratios across the C − 1 cut-offs theproportional odds assumption. As written above, a positive coefficient for a

regressor indicates that as values of the regressor increase so do the odds that

the response is greater than or equal to c. Although this is a natural way of

writing the model, because it means that for a positive β as x increases so

does the value of Y , it is not the only way of writing the model. In particular,

the model is sometimes written as

log

[

P ijc1 − P ijc

]

= γc + x′

ijβ + z′

ijTθj (c = 1, . . . , C − 1),

in which case the regression parameters β are identical but of opposite sign.

This alternate specification is commonly used in survival analysis models (see

section 6.3.2).

6.3.1 Partial Proportional Odds

As noted by Peterson and Harrell [77], violation of the proportional odds

assumption is not uncommon. Thus, they described a (fixed-effects) partial

proportional odds model in which covariates are allowed to have differential

6 Multilevel Models for Ordinal and Nominal Variables 247

effects on the C − 1 cumulative logits. Similarly, Terza [109] developed asimilar extension of the (fixed-effects) ordinal probit model. Hedeker and

Mermelstein [52, 53] utilize this extension within the context of a multilevel

ordinal regression model. For this, the model for the C − 1 cumulative logitscan be written as

log

[

P ijc1 − P ijc

]

= γc −[

(x∗ij)′βc + x

′

ijβ + z′

ijTθi]

(c = 1, . . . , C − 1),

where x∗ij is a h× 1 vector containing the values of observation ij on the setof h covariates for which proportional odds is not assumed. In this model, βcis a h × 1 vector of regression coefficients associated with these h covariates.Because βc carries the c subscript, the effects of these h covariates are allowed

to vary across the C − 1 cumulative logits. In many areas of research, thisextended model is useful. For example, suppose that in a alchohol reduction

study there are three response categories (abstinence, mild use, heavy use)

and suppose that an intervention designed to reduce drinking is not successful

in increasing the proportion of individuals in the abstinence category but is

successful in moving individuals from heavy to mild use. In this case, the

(covariate) effect of intervention group would not be observed on the first

cumulative logit, but would be observed on the second cumulative logit. This

extended model has been utilized in several articles [32, 114, 117], and a similar

Bayesian hierarchical model is described in Ishwaran [57].

In general, this extension of the proportional odds model is not problem-

atic, however, one caveat should be mentioned. For the explanatory variables

without proportional odds, the effects on the cumulative log odds, namely

(x∗ij)′βc, result in C − 1 non-parallel regression lines. These regression lines

inevitably cross for some values of x∗, leading to negative fitted values for the

response probabilities. For x∗ variables contrasting two levels of an explana-

tory variable (e.g., gender coded as 0 or 1), this crossing of regression lines

occurs outside the range of admissible values (i.e., < 0 or > 1). However, if the

explanatory variable is continuous, this crossing can occur within the range

of the data, and so, allowing for non-proportional odds can be problematic. A

solution to this dilemma is sometimes possible if the variable has, say, m levels

with a reasonable number of observations at each of these m levels. In this

case m − 1 dummy-coded variables can be created and substituted into themodel in place of the continuous variable. Alternatively, one might consider a

nominal response model using Helmert contrasts [15] for the outcome variable.

This approach, described in section 6.4, is akin to the sequential logit models

for nested or hierarchical response scales described in McCullagh and Nelder

[72].

248 Hedeker

6.3.2 Survival Analysis Models

Several authors have noted the connection between survival analysis models

and binary and ordinal regression models for survival data that are discrete

or grouped within time intervals (for practical introductions see Allison [6,

7], D’Agostino et al. [24], Singer and Willett [95]). This connection has been

utilized in the context of categorical multilevel or mixed-effects regression

models by many authors as well [42, 54, 94, 106, 108]. For this, assume that

time (of assessment) can take on only discrete positive values c = 1, 2, . . . , C.1

For each level-1 unit, observation continues until time Y ij at which point

either an event occurs (dij = 1) or the observation is censored (dij = 0),

where censoring indicates being observed at c but not at c+ 1. Define Pijc to

be the probability of failure, up to and including time interval c, that is,

Pijc = Pr(Y ij ≤ c),

and so the probability of survival beyond time interval c is simply 1 − Pijc.Because 1 − Pijc represents the survivor function, McCullagh [71] pro-

posed the following grouped-time version of the continuous-time proportional

hazards model

log[− log(1 − Pijc)] = γc + x′ijβ. (6.5)This is the aforementioned complementary log-log response function, which

can be re-expressed in terms of the cumulative failure probability, Pijc =

1 − exp(− exp(γc + x′ijβ)). In this model, xij includes covariates that varyeither at level 1 or 2, however they do not vary with time (i.e., they do not

vary across the ordered response categories). They may, however, represent

the average of a variable across time or the value of the covariate at the time

of the event.

The covariate effects in this model are identical to those in the grouped-

time version of the proportional hazards model described by Prentice and

Gloeckler [79]. As such, the β coefficients are also identical to the coefficients

in the underlying continuous-time proportional hazards model. Furthermore,

as noted by Allison [6], the regression coefficients of the model are invariant to

interval length. Augmenting the coefficients β, the threshold terms γc repre-

sent the logarithm of the integrated baseline hazard (i.e., when x = ∅). While

the above model is the same as that described in McCullagh [71], it is written

so that the covariate effects are of the same sign as the Cox proportional

hazards model. A positive coefficient for a regressor then reflects increasing

hazard (i.e., lower values of Y ) with greater values of the regressor. Adding

(standardized) random effects, we get

1 To make the connection to ordinal models more direct, time is here denoted as c,

however more commonly it is denoted as t in the survival analysis literature.

6 Multilevel Models for Ordinal and Nominal Variables 249

log[− log(1 − P ijc)] = γc + x′ijβ + z′ijTθj . (6.6)

This model is thus a multilevel ordinal regression model with a complementary

log-log response function instead of the logistic. Though the logistic model

has also been proposed for analysis of grouped and/or discrete time survival

data, its regression coefficients are not invariant to time interval length and it

requires the intervals to be of equal length [6]. As a result, the complementary

log-log response function is generally preferred.

In the ordinal treatment, survival time is represented by the ordered

outcome Y ij , which is designated as being censored or not. Alternatively,

each survival time can be represented as a set of dichotomous dummy codes

indicating whether or not the observation failed in each time interval that

was experienced [6, 24, 95]. Specifically, each survival time Y ij is represented

as a vector with all zeros except for its last element, which is equal to dij(i.e., = 0 if censored and = 1 for an event). The length of the vector for

observation ij equals the observed value of Y ij (assuming that the survival

times are coded as 1, 2, . . . , C). These multiple time indicators are then treated

as distinct observations in a dichotomous regression model. In a multilevel

model, a given cluster’s response vector Y j is then of size (∑nj

i=1 Y ij) × 1.This method has been called the pooling of repeated observations method

by Cupples et al. [23]. It is particularly useful for handling time-dependent

covariates and fitting non-proportional hazards models because the covariate

values can change across time. See Singer and Willett [96] for a detailed

treatment of this method.

For this dichotomous approach, define λijc to be the probability of failure

in time interval c, conditional on survival prior to c,

λijc = Pr(Y ij = c | Y ij ≥ c).

Similarly, 1 − λijc is the probability of survival beyond time interval c, con-ditional on survival prior to c. The multilevel proportional hazards model is

then written as

log[− log(1 − λijc)] = x′ijcβ + z′ijTθj , (6.7)

where now the covariates x can vary across time and so are denoted as xijc.

The first elements of x are usually timepoint dummy codes. Because the

covariate vector x now varies with c, this approach automatically allows for

time-dependent covariates, and relaxing the proportional hazards assumption

only involves including interactions of covariates with the timepoint dummy

codes.

Under the complementary log-log link function, the two approaches char-

acterized by (6.6) and (6.7) yield identical results for the parameters that do

not depend on c [28, 59]. Comparing these two approaches, notice that for

250 Hedeker

the ordinal approach each observation consists of only two pieces of data: the

(ordinal) time of the event and whether it was censored or not. Alternatively,

in the dichotomous approach each survival time is represented as a vector

of dichotomous indicators, where the size of the vector depends upon the

timing of the event (or censoring). Thus, the ordinal approach can be easier

to implement and offers savings in terms of the dataset size, especially as the

number of timepoints gets large, while the dichotomous approach is superior

in its treatment of time-dependent covariates and relaxing of the proportional

hazards assumption.

6.3.3 Estimation

For the ordinal models presented, the probability of a response in category c

for a given level-2 unit j, conditional on the random effects θ is equal to

Pr(Yij = c | θ) = Pijc − Pij,c−1 ,

where Pijc = 1/[1+exp(−ηijc)] under the logistic response function (formulasfor other response functions are given in section 6.2.3). Note that because

γ0 = −∞ and γC = ∞, Pij0 = 0 and PijC = 1. Here, ηijc denotes theresponse model, for example,

ηijc = γc −[

(x∗ij)′βc + x

′

ijβ + z′

ijTθi]

,

or one of the other variants of ηijc presented. In what follows, we’ll consider the

general model allowing for non-proportional odds, since the more restrictive

proportional odds model is just a special case (i.e., when βc = 0).

Let Yj denote the vector of ordinal responses from level-2 unit j (for the

nj level-1 units nested within). The probability of any pattern Yj conditional

on θ is equal to the product of the probabilities of the level-1 responses,

ℓ(Yj | θ) =nj∏

i=1

C∏

c=1

(Pijc − Pij,c−1)yijc , (6.8)

where yijc = 1 if Yij = c and 0 otherwise (i.e., for each ij-th observation,

yijc = 1 for only one of the C categories). For the ordinal representation of

the survival model, where right-censoring is present, the above likelihood is

generalized to

ℓ(Yj | θ) =nj∏

i=1

C∏

c=1

[

(Pijc − Pij,c−1)dij (1 − Pijc)1−dij]yijc

, (6.9)

where dij = 1 if Yij represents an event, or dij = 0 if Yij represents a censored

observation. Notice that (6.9) is equivalent to (6.8) when dij = 1 for all

6 Multilevel Models for Ordinal and Nominal Variables 251

observations. With right-censoring, because there is essentially one additional

response category (for those censored at the last category C), it is γC+1 = ∞and so Pij,C+1 = 1. In this case, parameters γc and βc with c = 1, . . . , C are

estimable, otherwise c only goes to C − 1.The marginal density of Y j in the population is expressed as the following

integral of the likelihood, ℓ(·), weighted by the prior density g(·),

h(Yj) =

∫

θ

ℓ(Yj | θ) g(θ) dθ, (6.10)

where g(θ) represents the multivariate standard normal density. The marginal

log-likelihood from the N level-2 units, logL =∑N

j log h(Yj), is then maxi-

mized to yield maximum likelihood estimates. For this, denote the conditional

likelihood as ℓj and the marginal density as hj . Differentiating first with

respect to the parameters that vary with c, let αk represent a particular

threshold γk or regression vector β∗

k, where k = 1, . . . , C if right-censoring

occurs, otherwise k = 1, . . . , C − 1. Then

∂ logL

∂αk=

N∑

j=1

h−1j∂hj∂αk

,

with

∂hj∂αk

=

∫

θ

nj∑

i=1

C∑

c=1

yijc

[

dij(∂Pijc)ack − (∂Pij,c−1)ac−1,k

Pijc − Pij,c−1

− (1 − dij)(∂Pijc)ack1 − Pijc

]

× ℓj g(θ)∂ηijk∂αk

dθ, (6.11)

where ∂ηijk/∂αk = 1 and −x∗ij for the thresholds and regression coefficients,respectively, and ack = 1 if c = k (and = 0 if c 6= k). Also, ∂Pijc represents thepdf of the response function; various forms of this are given in section 6.2.3.

For the parameters that do not vary with c, let ζ represent an arbitrary

parameter vector; then for β and the vector v(T ), which contains the unique

elements of the Cholesky factor T , we get

∂ logL

∂ζ=

N∑

j=1

h−1j

∫

θ

nj∑

i=1

C∑

c=1

yijc

[

dij∂Pijc − ∂Pij,c−1Pijc − Pij,c−1

− (1 − dij)∂Pijc

1 − Pijc

]

× ℓj g(θ)∂ηijc∂ζ

dθ, (6.12)

where∂ηijc∂β

= −xij ,∂ηijc

∂(v(T ))= −Jr(θ � zij),

252 Hedeker

and Jr is the elimination matrix of Magnus [69], which eliminates the elements

above the main diagonal. If T is an r × 1 vector of independent varianceterms (e.g., if zij is an r× 1 vector of level-1 or level-2 grouping variables, seesection 6.7), then ∂ηijc/∂T = zijθ in the equation above.

Fisher’s method of scoring can be used to provide the solution to these like-

lihood equations. For this, provisional estimates for the vector of parameters

Θ, on iteration ι are improved by

Θι+1 = Θι −{

E

[

∂2 logL

∂Θι ∂Θ′ι

]}−1∂ logL

∂Θι, (6.13)

where, following Bock and Lieberman [17], the information matrix, or minus

the expectation of the matrix of second derivatives, is given by

−E[

∂2 logL

∂Θι ∂Θ′ι

]

= E

N∑

j=1

h−2j∂hj∂Θι

(

∂hj∂Θι

)

′

.

Its estimator is obtained using the estimated parameter values and, at conver-

gence, the large-sample variance covariance matrix of the parameter estimates

is gotten as the inverse of the information matrix. The form on the right-hand

side of the above equation is sometimes called the “outer product of the

gradients.” It was proposed in the econometric literature by Berndt et al.

[12], and is often referred to as the BHHH method.

6.4 Multilevel Nominal Response Models

Let Y ij now denote a nominal variable associated with level-2 unit j and

level-1 unit i. Adding random effects to the fixed-effects multinomial logistic

regression model (see Agresti [2], Long [65]), we get that the probability that

Y ij = c (a response occurs in category c) for a given level-2 unit j is given by

pijc

= Pr(Y ij = c) =exp(η

ijc)

1 +∑C

h=2 exp(ηijh)for c = 2, 3, . . . , C, (6.14a)

pij1

= Pr(Y ij = 1) =1

1 +∑C

h=2 exp(ηijh), (6.14b)

where the multinomial logit ηijc

= x′ijβc + z′

ijTc θj . Comparing this to the

logit for ordered responses, we see that all of the covariate effects βc vary

across categories (c = 2, 3, . . . , C). Similarly for the random-effect variance

term Tc. As written above, an important distinction between the model for

ordinal and nominal responses is that the former uses cumulative comparisons

of the categories whereas the latter uses comparisons to a reference category.

6 Multilevel Models for Ordinal and Nominal Variables 253

This model generalizes Bock’s model for educational test data [14] by

including covariates xij , and by allowing a general random-effects design

vector zij including the possibility of multiple random effects θj . As discussed

by Bock [14], the model has a plausible interpretation. Namely, each nominal

category is assumed to be related to an underlying latent “response tendency”

for that category. The category c associated with the response variable Y ijis then the category for which the response tendency is maximal. Notice that

this assumption of C latent variables differs from the ordinal model where only

one underlying latent variable is assumed. Bock [15] refers to the former as

the extremal concept and the latter as the aforementioned threshold concept,

and notes that both were introduced into psychophysics by Thurstone [111].

The two are equivalent only for the dichotomous case (i.e., when there are

only two response categories).

The model as written above allows estimation of any pairwise comparisons

among the C response categories. As characterized in Bock [14], it is benefical

to write the nominal model to allow for any possible set of C − 1 contrasts.For this, the category probabilities are written as

pijc

=exp(η

ijc)

∑Ch=1 exp(ηijh)

for c = 1, 2, . . . , C, (6.15)

where now

ηijc

= x′ijΓdc + (z′

ij � θ′

j)J′

r∗Λdc . (6.16)

Here, D is the (C − 1) × C matrix containing the contrast coefficients forthe C − 1 contrasts between the C logits and dc is the cth column vectorof this matrix. The s × (C − 1) parameter matrix Γ contains the regressioncoefficients associated with the s covariates for each of the C − 1 contrasts.Similarly, Λ contains the random-effect variance parameters for each of the

C − 1 contrasts. Specifically,

Λ = [ v(T1) v(T2) . . . v(TC−1) ] ,

where v(Tc) is the r∗ × 1 vector (r∗ = r(r + 1)/2) of elements below and on

the diagonal of the Cholesky (lower-triangular) factor Tc, and Jr∗ is the afore-

mentioned elimination matrix of Magnus [69]. This latter matrix is necessary

to ensure that the appropriate terms from the 1×r2 vector resulting from theKronecker product (z′ij � θ

′

j) are multiplied with the r∗ × 1 vector resulting

from Λdc. For the case of a random-intercepts model, the model simplifies to

ηijc

= x′ijΓdc + Λdc θj ,

with Λ as the 1 × (C − 1) vector Λ = [ σ1 σ2 . . . σC−1 ].Notice that if D equals

254 Hedeker

D =

0 1 0 . . . 0

0 0 1 . . . 0

. . . . . . .

0 0 0 . . . 1

,

the model simplifies to the earlier representation in (6.14a) and (6.14b). The

current formulation, however, allows for a great deal of flexibility in the types

of comparisons across the C response categories. For example, if the categories

are ordered, an alternative to the cumulative logit model of the previous

section is to employ Helmert contrasts [15] within the nominal model. For

this, with C = 4, the contrast matrix would be

D =

−1 13 13 130 −1 12 120 0 −1 1

.

Helmert contrasts are similar to the category comparisons of continuation-

ratio logit models, as described within a mixed model formulation by Ten Have

and Uttal [108]. However, the Helmert contrasts above are applied to the

category logits, rather then the category probabilities as in continuation-ratio

models.

6.4.1 Parameter Estimation

Estimation follows the procedure described for ordinal outcomes. Specifically,

letting Yj denote the vector of nominal responses from level-2 unit j (for the

nj level-1 units nested within), the probability of any Yj conditional on the

random effects θ is equal to the product of the probabilities of the level-1

responses

ℓ(Yj | θ) =nj∏

i=1

C∏

c=1

(pijc)yijc , (6.17)

where yijc = 1 if Yij = c, and 0 otherwise. The marginal density of the

response vector Yj is again given by (6.10). The marginal log-likelihood from

the N level-2 units, logL =∑N

j log h(Yj), is maximized to obtain maximum

likelihood estimates of Γ and Λ. Specifically, using ∆ to represent either

parameter matrix,

∂ logL

∂∆′=

N∑

j=1

h−1(Yj)

∫

θ

[

nj∑

i=1

D (yij − Pij) � ∂∆]

× ℓ(Yj | θ) g(θ) dθ, (6.18)

where

6 Multilevel Models for Ordinal and Nominal Variables 255

∂Γ = x′ij , ∂Λ = [Jr∗(θ � zij)]′

,

yij is the C × 1 indicator vector, and Pij is the C × 1 vector obtained byapplying (6.15) for each category. As in the ordinal case, Fisher’s method of

scoring can be used to provide the solution to these likelihood equations.

6.5 Computational Issues

In order to solve the above likelihood solutions for both the ordinal and

nominal models, integration over the random-effects distribution must be

performed. Additionally, the above likelihood solutions are only in terms of

the regression parameters and variance-covariance parameters of the random-

effects distribution. Often, estimation of the random effects is also of interest.

These issues are described in great detail in Skrondal and Rabe-Hesketh [98];

here, we discuss some of the relevant points.

6.5.1 Integration over θ

Various approximations for evaluating the integral over the random-effects

distribution have been proposed in the literature; several of these are com-

pared in chapter 9. Perhaps the most frequently used methods are based

on first- or second-order Taylor expansions. Marginal quasi-likelihood (MQL)

involves expansion around the fixed part of the model, whereas penalized or

predictive quasi-likelihood (PQL) additionally includes the random part in its

expansion [39]. Both of these are available in the MLwiN software program

[84]. Unfortunately, several authors [19, 87, 90] have reported downwardly

biased estimates using these procedures in certain situations, especially for

the first-order expansions.

Raudenbush et al. [87] proposed an approach that uses a combination of

a fully multivariate Taylor expansion and a Laplace approximation. Based on

the results in Raudenbush et al. [87], this method yields accurate results and is

computationally fast. Also, as opposed to the MQL and PQL approximations,

the deviance obtained from this approximation can be used for likelihood-ratio

tests. This approach has been incorporated into the HLM software program

[86].

Numerical integration can also be used to perform the integration over the

random-effects distribution. Specifically, if the assumed distribution is normal,

Gauss-Hermite quadrature can approximate the above integral to any practi-

cal degree of accuracy [104]. Additionally, like the Laplace approximation, the

numerical quadrature approach yields a deviance that can be readily used for

likelihood-ratio tests. The integration is approximated by a summation on a

specified number of quadrature points Q for each dimension of the integration.

256 Hedeker

The solution via quadrature can involve summation over a large number of

points, especially as the number of random effects is increased. For example,

if there is only one random effect, the quadrature solution requires only one

additional summation over Q points relative to the fixed effects solution. For

models with r > 1 random effects, however, the quadrature is performed

over Qr points, and so becomes computationally burdensome for r > 5 or

so. Also, Lesaffre and Spiessens [61] present an example where the method

only gives valid results for a high number of quadrature points. These authors

advise practitioners to routinely examine results for the dependence on Q.

To address these issues, several authors have described a method of adaptive

quadrature that uses relatively few points per dimension (e.g., 3 or so), which

are adapted to the location and dispersion of the distribution to be integrated

[18, 64, 78, 80]. Simulations show that adaptive quadrature performs well

in a wide variety of situations and typically outperforms ordinary quadrature

[82]. Several software packages have implemented ordinary or adaptive Gauss-

Hermite quadrature, including EGRET [22], GLLAMM [81], LIMDEP [40],

MIXOR [49], MIXNO [46], Stata [101], and SAS PROC NLMIXED [93].

Another approach that is commonly used in econometrics and transporta-

tion research uses simulation methods to integrate over the random-effects

distribution (see the introductory overview by Stern [102] and the excellent

book by Train [112]). When used in conjunction with maximum likelihood

estimation, it is called “maximum simulated likelihood” or “simulated maxi-

mum likelihood.” The idea behind this approach is to draw a number of values

from the random-effects distribution, calculate the likelihood for each of these

draws, and average over the draws to obtain a solution. Thus, this method

maximizes a simulated sample likelihood instead of an exact likelihood, but

can be considerably faster than quadrature methods, especially as the number

of random effects increases [41]. It is a very flexible and intuitive approach

with many potential applications (see Drukker [27]). In particular, Bhat [13]

and Glasgow [36] describe this estimation approach for multilevel models

of nominal outcomes. In terms of software, LIMDEP [40] has included this

estimation approach for several types of outcome variables, including nominal

and ordinal, and Haan and Uhlendorff [41] describe a Stata routine for nominal

data.

Bayesian approaches, such as the use of Gibbs sampling [33] and related

methods [105], can also be used to integrate over the random effects distribu-

tion. This approach is described in detail in chapter 2. For nominal responses,

Daniels and Gatsonis [25] use this approach in their multilevel polychotomous

regression model. Similarly, Ishwaran [57] utilize Bayesian methods in model-

ing multilevel ordinal data. The freeware BUGS software program [100] can be

used to facilitate estimation via Gibbs sampling. In this regard, Marshall and

Spiegelhalter [70] provides an example of multilevel modeling using BUGS,

including some syntax and discussion of the program.

6 Multilevel Models for Ordinal and Nominal Variables 257

6.5.2 Estimation of Random Effects and Probabilities

In many cases, it is useful to obtain estimates of the random effects and also to

obtain fitted marginal probabilities. The random effects θj can be estimated

using empirical Bayes methods [16]. For the univariate case, this estimator θ̂jis given by the mean of the posterior distribution,

θ̂j = E (θj | Yj) =1

h(Yj)

∫

θ

θj ℓ(·) g(θ) dθ, (6.19)

where ℓ(·) is the conditional likelihood for the particular model (i.e., ordinalor nominal). The variance of the posterior distribution is obtained as

Var(θ̂j | Yj) =1

h(Yj)

∫

θ

(θj − θ̂j)2 ℓ(·) g(θ) dθ.

These quantities may be used, for example, to evaluate the response proba-

bilities for particular level-2 units (e.g., person-specific trend estimates).

To obtain estimated marginal probabilities (e.g., the estimated response

probabilities of the control group across time), an additional step is required

for models with non-linear response functions (e.g., the models considered

in this paper). First, so-called “subject-specific” probabilities [75, 118] are

estimated for specific values of covariates and random effects, say θ∗. These

subject-specific estimates indicate, for example, the response probability for

a subject with random effect level θ∗ in the control group at a particu-

lar timepoint. Denoting these subject-specific probabilities as P̂ss, marginal

probabilities P̂m can then be obtained by numerical quadrature, namely

P̂m =∫

θP̂ss g(θ) dθ, or by marginalizing the scale of the regression coef-

ficients [51, p. 179]. Continuing with our example, the marginalized estimate

would indicate the estimated response probability for the entire control group

at a particular timepoint. Both subject-specific and marginal estimates have

their uses, since they are estimating different quantities, and several authors

have characterized the differences between the two [45, 62, 75].

6.6 Intraclass Correlation

For a random-intercepts model (i.e., zj = 1nj ) it is often of interest to express

the level-2 variance in terms of an intraclass correlation. For this, one can make

reference to the threshold concept and the underlying latent response tendency

that determines the observed response. For the ordinal logistic model assum-

ing normally distributed random-effects, the estimated intraclass correlation

equals σ̂2/(σ̂2 + π2/3), where the latter term in the denominator represents

the variance of the underlying latent response tendency. As mentioned earlier,

for the logistic model, this variable is assumed to be distributed as a standard

258 Hedeker

logistic distribution with variance equal to π2/3. For a probit model this term

is replaced by 1, the variance of the standard normal distribution.

For the nominal model, one can make reference to multiple underlying

latent response tendencies, denoted as yijc

, and the associated regression

model including level-1 residuals ǫijc

yijc

= x′ijβc + z′

ijTc θj + ǫijc c = 1, 2, . . . , C.

As mentioned earlier, for a particular ij-th unit, the category c associated

with the nominal response variable Y ij is the one for which the latent yijcis maximal. Since, in the common reference cell formulation, c = 1 is the

reference category, T1 = β1 = 0, and so the model can be rewritten as

yijc

= x′ijβc + z′

ijTc θj + (ǫijc − ǫij1) c = 2, . . . , C,

for the latent response tendency of category c relative to the reference cat-

egory. It can be shown that the level-1 residuals ǫijc for each category are

distributed according to a type I extreme-value distribution [see 68, p. 60].

It can further be shown that the standard logistic distribution is obtained

as the difference of two independent type I extreme-value variates [see 72,

pp. 20 and 142]. As a result, the level-1 variance is given by π2/3, which

is the variance for a standard logistic distribution. The estimated intraclass

correlations are thus calculated as rc = σ̂2c/(σ̂

2c + π

2/3), where σ̂2c is the

estimated level-2 variance assuming normally-distributed random intercepts.

Notice that C − 1 intraclass correlations are estimated, one for each categoryc versus the reference category. As such, the cluster influence on the level-1

responses is allowed to vary across the nominal response categories.

6.7 Heterogeneous Variance Terms

Allowing for separate random-effect variance terms for groups of either i or j

units is sometimes important. For example, in a twin study it is often necessary

to allow the intra-twin correlation to differ between monozygotic and dizygotic

twins. In this situation, subjects (i = 1, 2) are nested within twin pairs (j =

1, . . . , N). To allow the level-2 variance to vary for these two twin-pair types,

the random-effects design vector zij is specified as a 2 × 1 vector of dummycodes indicating monozygotic and dizygotic twin pair status, respectively. T

(or Tc in the nominal model) is then a 2 × 1 vector of independent random-effect standard deviations for monozygotics and dizygotics, and the cluster

effect θj is a scalar that is pre-multiplied by the vector T . For example, for a

random-intercepts proportional odds model, we would have

log

[

P ijc1 − P ijc

]

= γc −{

x′ijβ + [MZ j DZ j ]

[

σδ (MZ )σδ (DZ )

]

θj

}

,

6 Multilevel Models for Ordinal and Nominal Variables 259

where MZ j and DZ j are dummy codes indicating twin pair status (i.e., if

MZ j = 1 then DZ j = 0, and vice versa).

Notice, that if the probit formulation is used and the model has no covari-

ates (i.e., only an intercept, xij = 1), the resulting intraclass correlations

ICCMZ =σ2

δ (MZ )

σ2δ (MZ ) + 1

and ICCDZ =σ2

δ (DZ )

σ2δ (DZ ) + 1

are polychoric correlations (for ordinal responses) or tetrachoric correlations

(for binary responses) for the within twin-pair data. Adding covariates then

yields adjusted tetrachoric and polychoric correlations. Because estimation of

polychoric and tetrachoric correlations is often important in twin and genetic

studies, these models are typically formulated in terms of the probit link.

Comparing models that allow homogeneous versus heterogeneous subgroup

random-effects variance, thus allows testing of whether the tetrachoric (or

polychoric) correlations are equal across the subgroups.

The use of heterogeneous variance terms can also be found in some item

response theory (IRT) models in the educational testing literature [14, 16,

92]. Here, item responses (i = 1, 2, . . . ,m) are nested within subjects (j =

1, 2, . . . , N) and a separate random-effect standard deviation (i.e., an element

of the m×1 vector T ) is estimated for each test item (i.e., each i unit). In themultilevel model this is accomplished by specifying zij as an m× 1 vector ofdummy codes indicating the repeated items. To see this, consider the popular

two-parameter logistic model for dichotomous responses [66] that specifies the

probability of a correct response to item i (Y ij = 1) as a function of the ability

of subject j (θj),

Pr(Y ij = 1) =1

1 + exp[−ai(θj − bi)],

where ai is the slope parameter for item i (i.e., item discrimination), and biis the threshold or difficulty parameter for item i (i.e., item difficulty). The

distribution of ability in the population of subjects is assumed to be normal

with mean 0 and variance 1 (i.e., the usual assumption for the random effects

θj in the multilevel model). As noted by Bock and Aitkin [16], it is convenient

to let ci = −aibi and write

Pr(Y ij = 1) =1

1 + exp[−(ci + aiθj)],

which can be recast in terms of the logit of the response as

logitij = log

[

pij

1 − pij

]

= ci + aiθj .

260 Hedeker

As an example, suppose that there are four items. This model can be repre-

sented in matrix form as

logit1jlogit2jlogit3jlogit4j

=

1 0 0 0

0 1 0 0

0 0 1 0

0 0 0 1

Xj

c1c2c3c4

c

+

1 0 0 0

0 1 0 0

0 0 1 0

0 0 0 1

Zj

a1a2a3a4

a

( θj),

showing that this IRT model is a multilevel model that allows the random

effect variance terms to vary across items (level-1). The usual IRT notation

is a bit different than the multilevel notation, but c simply represents the

fixed-effects (i.e., β) and a is the the random-effects standard deviation vector

T ′ = [ σδ1 σδ2 σδ3 σδ4 ].

The elements of the T vector can also be viewed as the (unscaled) factor

loadings of the items on the (unidimensional) underlying ability variable (θ).

A simpler IRT model that constrains these factor loadings to be equal is the

one-parameter logistic model, the so-called Rasch model [116]. This constraint

is achieved by setting Zj = 1nj and a = a in the above model. Thus, the

Rasch model is simply a random-intercepts logistic regression model with

item indicators for X.

Unlike traditional IRT models, the multilevel formulation of the model

easily allows multiple covariates at either level (i.e., items or subjects). This

and other advantages of casting IRT models as multilevel models are described

in detail by Adams et al. [1] and Rijmen et al. [89]. In particular, this allows a

model for examining whether item parameters vary by subject characteristics,

and also for estimating ability in the presence of such item by subject in-

teractions. Interactions between item parameters and subject characteristics,

often termed item bias [20], is an area of active psychometric research. Also,

although the above illustration is in terms of a dichotomous response model,

the analogous multilevel ordinal and nominal models apply. For ordinal items

responses, application of the cumulative logit multilevel models yields what

Thissen and Steinberg [110] have termed “difference models,” namely, the

treatment of ordinal responses as developed by Samejima [92] within the

IRT-context. Similarly, in terms of nominal responses, the multilevel model

yields the nominal IRT model developed by Bock [14].

6.8 Health Services Research Example

The McKinney Homeless Research Project (MHRP) study [55, 56] in San

Diego, CA was designed to evaluate the effectiveness of using section 8 certifi-

cates as a means of providing independent housing to the severely mentally ill

homeless. Section 8 housing certificates were provided from the Department

6 Multilevel Models for Ordinal and Nominal Variables 261

of Housing and Urban Development (HUD) to local housing authorities in

San Diego. These housing certificates, which require clients to pay 30% of

their income toward rent, are designed to make it possible for low income in-

dividuals to choose and obtain independent housing in the community. Three

hundred sixty-one clients took part in this longitudinal study employing a

randomized factorial design. Clients were randomly assigned to one of two

types of supportive case management (comprehensive vs. traditional) and to

one of two levels of access to independent housing (using section 8 certificates).

Eligibility for the project was restricted to individuals diagnosed with a severe

and persistent mental illness who were either homeless or at high risk of

becoming homeless at the start of the study. Individuals’ housing status was

classified at baseline and at 6, 12, and 24 month follow-ups.

In this illustration, focus will be on examining the effect of access to

section 8 certificates on repeated housing outcomes across time. Specifi-

cally, at each timepoint each subjects’ housing status was classified as either

streets/shelters, community housing, or independent housing. This outcome

can be thought of as ordinal with increasing categories indicating improved

housing outcomes. The observed sample sizes and response proportions for

these three outcome categories by group are presented in Table 6.1.

Table 6.1. Housing status across time by group: response proportions and sample

sizes.

timepoint

group status baseline 6-months 12-months 24-months

control street .555 .186 .089 .124

community .339 .578 .582 .455

independent .106 .236 .329 .421

n 180 161 146 145

section 8 street .442 .093 .121 .120

community .414 .280 .146 .228

independent .144 .627 .732 .652

n 181 161 157 158

These observed proportions indicate a general decrease in street living and

an increase in independent living across time for both groups. The increase

in independent housing, however, appears to occur sooner for the section 8

group relative to the control group. Regarding community living, across time

this increases for the control group and decreases for the section 8 group.

262 Hedeker

There is some attrition across time; attrition rates of 19.4% and 12.7%

are observed at the final timepoint for the control and section 8 groups, re-

spectively. Since estimation of model parameters is based on a full-likelihood

approach, the missing data are assumed to be “ignorable” conditional on

both the model covariates and the observed responses [60]. In longitudinal

studies, ignorable nonresponse falls under the “missing at random” (MAR)

assumption introduced by Rubin [91], in which the missingness depends only

on observed data. In what follows, since the focus is on describing application

of the various multilevel regression models, we will make the MAR assumption.

A further approach, however, that does not rely on the MAR assumption (e.g.,

a multilevel pattern-mixture model as described in Hedeker and Gibbons [50])

could be used. Missing data issues are described more fully in chapter 10.

6.8.1 Ordinal Response Models

To prepare for the ordinal analyses, the observed cumulative logits across time

for the two groups are plotted in Figures 6.3 and 6.4 The first cumulative

logit compares independent and community housing versus street living (i.e.,

categories 2 & 3 combined versus 1), while the second cumulative logit com-

pares independent housing versus community housing and street living (i.e.,

category 3 versus 2 and 1 combined). For the proportional odds model to hold,

these two plots should look the same, with the only difference being the scale

difference on the y-axis. As can be seen, these plots do not look that similar.

For example, the post-baseline group differences do not appear to be the

same for the two cumulative logits. In particular, it appears that the section 8

group does better more consistently in terms of the second cumulative logit

(i.e., independent versus community and street housing). This would imply

that the proportional odds model is not reasonable for these data.

To assess this more rigorously, two ordinal logistic multilevel models were

fit to these data, the first assuming a proportional odds model and the sec-

ond relaxing this assumption. For both analyses, the repeated housing status

classifications were modeled in terms of time effects (6, 12, and 24 month

follow-ups compared to baseline), a group effect (section 8 versus control), and

group by time interaction terms. The first analysis assumes these effects are

the same across the two cumulative logits of the model, whereas the second

analysis estimates effects for each explanatory variable on each of the two

cumulative logits. In terms of the multilevel part of the model, only a random

subject effect was included in both analyses. Results from these analyses are

given in Table 6.2.

The proportional odds model indicates significant time effects for all time-

points relative to baseline, but only significant group by time interactions for

the 6 and 12 month follow-ups. Marginally significant effects are obtained for

the section 8 effect and the section 8 by t3 (24-months) interaction. Thus,

6 Multilevel Models for Ordinal and Nominal Variables 263

−0.5

0.0

0.5

1.0

1.5

2.0

2.5

Em

piric

allo

gits

0 6 12 24

Months since baseline

ControlSection 8

Fig. 6.3. First cumulative logit values across time by group.

−2

−1.5

−1.0

−0.5

0.0

0.5

1.0

Em

piric

allo

gits

0 6 12 24

Months since baseline

ControlSection 8

Fig. 6.4. Second cumulative logit values across time by group.

the analysis indicates that the control group moves away from street living

to independent living across time, and that this improvement is more pro-

nounced for section 8 subjects at the 6 and 12 month follow-up. Because the

section 8 by t3 interaction is only marginally significant, the groups do not

differ significantly in housing status at the 24-month follow-up as compared

to baseline.

However, comparing log-likelihood values clearly rejects the proportional

odds assumption (likelihood ratio χ27 = 52.14) indicating that the effects of

the explanatory variables cannot be assumed identical across the two cumu-

264 Hedeker

Table 6.2. Housing status across time: Ordinal logistic model estimates and stan-

dard errors (se).

Proportional Odds Non-Proportional Odds

Non-street1 Independent2

term estimate se estimate se estimate se

intercept −.220 .203 −.322 .218

threshold 2.744 .110 2.377 .279

t1 (6 month vs base) 1.736 .233 2.297 .298 1.079 .358

t2 (12 month vs base) 2.315 .268 3.345 .450 1.645 .336

t3 (24 month vs base) 2.499 .247 2.821 .369 2.145 .339

section 8 (yes=1, no=0) .497 .280 .592 .305 .323 .401

section 8 by t1 1.408 .334 .566 .478 2.023 .478

section 8 by t2 1.173 .360 −.958 .582 2.016 .466

section 8 by t3 .638 .331 −.366 .506 1.073 .472

subject sd 1.459 .106 1.457 .112

−2 log L 2274.39 2222.25

bold indicates p < .05, italic indicates .05 < p < .101 logit comparing independent and community housing vs. street2 logit comparing independent housing vs. community housing and street

lative logits. Interestingly, none of the section 8 by time interaction terms are

significant in terms of the non-street logit (i.e., comparing categories 2 and 3

versus 1), while all of them are significant in terms of the independent logit

(i.e., comparing category 3 versus 1 and 2 combined). Thus, as compared to

baseline, section 8 subjects are more likely to be in independent housing at

all follow-up timepoints, relative to the control group.

In terms of the random subject effect, it is clear that the data are corre-

lated within subjects. Expressed as an intra-class correlation, the attributable

variance at the subject-level equals .39 for both models. Also, the Wald test is

highly significant in terms of rejecting the null hypothesis that the (subject)

population standard deviation equals zero. Strictly speaking, as noted by

Raudenbush and Bryk [85] and others, this test is not to be relied upon,

especially as the population variance is close to zero. In the present case, the

actual significance test is not critical because it is more or less assumed that

the population distribution of the subject effects will not have zero variance.

6 Multilevel Models for Ordinal and Nominal Variables 265

6.8.2 Nominal Response Models

For the initial set of analyses with nominal models, reference category con-

trasts were used and street/shelter was chosen as the reference category. Thus,

the first comparison compares community to street responses, and the second

compares independent to street responses. A second analysis using Helmert

contrasts will be described later.

Corresponding observed logits for the reference-cell comparisons by group

and time are given in Figures 6.5 and 6.6. Comparing these plots, different

patterns for the post-baseline group differences are suggested. It seems that

the non-section 8 group does better in terms of the community versus street

comparison, whereas the section 8 group is improved for the independent

versus street comparison. Further, the group differences appear to vary across

time. The subsequent analyses will examine these visual impressions of the

data.

−0.5

0.0

0.5

1.0

1.5

2.0

Em

piric

allo

gits

0 6 12 24

Months since baseline

ControlSection 8

Fig. 6.5. First reference-cell logit values across time by group.

To examine the sensivity of the results to the normality assumption for

the random effects, two multilevel nominal logistic regression models were

fit to these data assuming the random effects were normally and uniformly

distributed, respectively. Tables 6.3 and 6.4 list results for the two response

category comparisons of community versus street and independent versus

street, respectively. The time and group effects are the same as in the previous

ordinal analyses.

The results are very similar for the two multilevel models. Thus, the

random-effects distributional form does not seem to play an important role

for these data. Subjects in the control group increase both independent and

community housing relative to street housing at all three follow-ups, as com-

266 Hedeker

−2.0

−1.5

−1.0

−0.5

0.0

0.5

1.0

1.5

2.0

Em

piric

allo

gits

0 6 12 24

Months since baseline

ControlSection 8

Fig. 6.6. Second reference-cell logit values across time by group.

Table 6.3. Housing status across time: Nominal model estimates and standard

errors (se).

Community versus Street

Normal prior Uniform prior

term estimate se estimate se

intercept −.452 .192 −.473 .184

t1 (6 month vs base) 1.942 .312 1.850 .309

t2 (12 month vs base) 2.820 .466 2.686 .457

t3 (24 month vs base) 2.259 .378 2.143 .375

section 8 (yes=1, no=0) .521 .268 .471 .258

section 8 by t1 −.135 .490 −.220 .484

section 8 by t2 −1.917 .611 −1.938 .600

section 8 by t3 −.952 .535 −.987 .527

subject sd .871 .138 .153 .031

−2 log L 2218.73 2224.74

bold indicates p < .05, italic indicates .05 < p < .10

6 Multilevel Models for Ordinal and Nominal Variables 267

Table 6.4. Housing status across time: Nominal model estimates and standard

errors (se).

Independent versus Street

Normal prior Uniform prior

term estimate se estimate se

intercept −2.675 .367 −2.727 .351

t1 (6 month vs base) 2.682 .425 2.540 .422

t2 (12 month vs base) 4.088 .559 3.916 .551

t3 (24 month vs base) 4.099 .469 3.973 .462

section 8 (yes=1, no=0) .781 .491 .675 .460

section 8 by t1 2.003 .614 2.016 .605

section 8 by t2 .548 .694 .645 .676

section 8 by t3 .304 .615 .334 .600

subject sd 2.334 .196 .490 .040

−2 log L 2218.73 2224.74

bold indicates p < .05, italic indicates .05 < p < .10

pared to baseline. Compared to controls, the increase in community versus

street housing is less pronounced for section 8 subjects at 12 months, but

not statistically different at 6 months and only marginally different at 24

months. Conversely, as compared to controls, the increase in independent

versus street housing is more pronounced for section 8 subjects at 6 months,

but not statistically different at 12 or 24 months. Thus, both groups reduce

the degree of street housing, but do so in somewhat different ways. The control

group subjects are shifted more towards community housing, whereas section 8

subjects are more quickly shifted towards independent housing.

As in the ordinal case, the Wald tests are all significant for the inclusion of

the random effects variance terms. A likelihood-ratio test also clearly supports

inclusion of the random subject effect (likelihood ratio χ22 = 134.3 and 128.3

for the normal and uniform distribution, respectively, as compared to the

fixed-effects model, not shown). Expressed as intraclass correlations, r1 = .19

and r2 = .62 for community versus street and independent versus street,

respectively. Thus, the subject influence is much more pronounced in terms of

distinguishing independent versus street living, relative to community versus

street living. This is borne out by contrasting models with separate versus a

common random-effect variance across the two category contrasts (not shown)

which yields a highly significant likelihood ratio χ21 = 49.2 favoring the model

with separate variance terms.

An analysis was also done to examine if the random-effect variance terms

varied significantly by treatment group. The deviance (−2 logL) for this

268 Hedeker

model, assuming normally distributed random effects, equaled 2218.43, which

was nearly identical to the value of 2218.73 (from Tables 6.3 and 6.4) for

the model assuming homogeneous variances across groups. The control group

and section 8 group estimates of the subject standard deviations were respec-

tively .771 (se = .182) and .966 (se = .214) for the community versus street

comparison, and 2.228 (se = .299) and 2.432 (se = .266) for the indepen-

dent versus street comparison. Thus, the homogeneity of variance assumption

across treatment groups is clearly not rejected.

Finally, Table 6.5 lists the results obtained for an analysis assuming

normally-distributed random effects and using Helmert contrasts for the three

response categories. From this analysis, it is interesting that none of the sec-

tion 8 by time interaction terms are observed to be statistically significant for

the first Helmert contrast (i.e., comparing street to non-street housing). Thus,

group assignment is not significantly related to housing when considering sim-

ply street versus non-street housing outcomes. However, the second Helmert

contrast that contrasts the two types of non-street housing (i.e., independent

versus community) does reveal the benefical effect of the section 8 certificate

in terms of the positive group by time interaction terms. Again, the section 8

group is more associated with independent housing, relative to community

housing, than the non-section 8 group. In many ways, the Helmert contrasts,

with their intuitive interpretations, represent the best choice for the analysis

of these data.

Table 6.5. Housing status across time: Nominal model estimates and standard

errors (se) using Helmert contrasts.

Independent

& Community Independent

vs Street vs Community

term estimate se estimate se

intercept −1.042 .163 −1.112 .163

t1 (6 month vs base) 1.541 .215 .371 .187

t2 (12 month vs base) 2.303 .323 .634 .176

t3 (24 month vs base) 2.119 .258 .920 .179

section 8 (yes=1, no=0) .434 .222 .130 .213

section 8 by t1 .623 .330 1.069 .253

section 8 by t2 −.457 .401 1.233 .256

section 8 by t3 −.216 .345 .628 .255

subject sd 1.068 .099 .732 .083

−2 log L = 2218.73

bold indicates p < .05, italic indicates .05 < p < .10

6 Multilevel Models for Ordinal and Nominal Variables 269

6.9 Discussion

Multilevel ordinal and multinomial logistic regression models are described

for the analysis of categorical data. These models are useful for analysis of

outcomes with more than two response categories. By and large, the models

are seen as extensions of the multilevel logistic regression model. However,

they generalize the model in different ways. The ordinal model uses cumulative

dichotomizations of the categorical outcome. Alternatively, the nominal model

typically uses dichotomizations that are based on selecting one category as the

reference that the others are each compared to. This chapter has also described

how other comparisons can be embedded within the nominal model.

For ordinal data, both proportional odds and non-proportional odds mod-

els are considered. Since, as noted by Peterson and Harrell [77], examples of

non-proportional odds are not difficult to find, the latter model is especially

attractive for analyzing ordinal outcomes. In the example presented, the non-

proportional odds model provided more specific information about the effect

of section 8 certificates. Namely, as compared to baseline, these certificates

were effective in increasing independent housing (versus community housing

and street living combined) at all follow-up timepoints. Interestingly, the same

could not be said when comparing independent and community housing com-

bined versus street living. Thus, the use of the non-proportional odds model

was helpful in elucidating a more focused analysis of the effect of the section 8

program.

For the nominal model, both reference cell and Helmert contrasts were

applied in the analysis of these data. The former indicated an increase for

community relative to street housing for the non-section 8 group, and an

increase for independent relative to street housing for the section 8 group.

Alternatively, the Helmert contrasts indicated that the groups did not differ

in terms of non-street versus street housing, but did differ in terms of the

type of non-street housing (i.e., the section 8 group was more associated with

independent housing). In either case, the nominal model makes an assump-

tion that has been referred to as “independence of irrelevant alternatives”

[10, 67, 68]. This is because the effect of an explanatory variable comparing two

categories is the same regardless of the total number of categories considered.

This assumption is generally reasonable when the categories are distinct and

dissimilar, and unreasonable as the nominal categories are seen as substitutes

for one another [8, 73]. Furthermore, McFadden [74] notes that the multino-

mial logistic regression model is relatively robust in many cases in which this

assumption is implausible. In the present example, the outcome categories are

fairly distinct and so the assumption would seem to be reasonable for these

data. The possibility of relaxing this assumption, though, for a more general

multilevel nominal regression model is discussed in detail in Train [112].

270 Hedeker

The example presented illustrated the usefulness of the multilevel approach

for longitudinal categorical data. In particular, it showed the many possible

models and category comparisons that are possible if the response variable

has more than two categories. In terms of the multilevel part of the model,

only random-intercepts models were considered in the data analysis. However,

in describing model development, multiple random effects were allowed. An

analysis of these data incorporating random subject intercepts and linear

trends is discussed in Hedeker [46]. Additionally, the data had a relatively

simple multilevel structure, in that there were only two levels, namely, re-

peated observations nested within subjects. Extensions of both the ordinal

and nominal models for three and higher level is possible in the MLwiN [84]

and HLM [86] software programs.

Acknowledgement. Thanks are due to Drs. Richard Hough and Michael Hurlburt

for use of the longitudinal data, to Jan de Leeuw for organizing this collection,

and especially to Erik Meijer for many helpful comments on previous drafts of this

chapter. This work was supported by National Institutes of Mental Health Grant

MH56146.

References

1. R. J. Adams, M. Wilson, and M. Wu. Multilevel item response models: An ap-

proach to errors in variable regression. Journal of Educational and Behavioral

Statistics, 22:47–76, 1997.

2. A. Agresti. Categorical Data Analysis. Wiley, New York, 1990.

3. A. Agresti, J. G. Booth, J. P. Hobart, and B. Caffo. Random-effects modeling

of categorical response data. Sociological Methodology, 30:27–80, 2000.

4. A. Agresti and J. B. Lang. A proportional odds model with subject-specific

effects for repeated ordered categorical responses. Biometrika, 80:527–534,

1993.

5. A. Agresti and R. Natarajan. Modeling clustered ordered categorical data: A

survey. International Statistical Review, 69:345–371, 2001.

6. P. D. Allison. Discrete-time methods for the analysis of event histories. Soci-

ological Methodology, 13:61–98, 1982.

7. P. D. Allison. Survival Analysis using the SAS System: A Practical Guide. SAS

Institute, Cary, NC, 1995.

8. T. Amemiya. Qualitative response models: A survey. Journal of Economic

Literature, 19:483–536, 1981.

9. D. A. Anderson and M. Aitkin. Variance component models with binary

response: Interviewer variability. Journal of the Royal Statistical Society, Se-

ries B, 47:203–210, 1985.

10. K. J. Arrow. Social Choice and Individual Values. Wiley, New York, 1951.

11. D. J. Bartholomew and M. Knott. Latent Variable Models and Factor Analysis.

Oxford University Press, New York, 2nd edition, 1999.

6 Multilevel Models for Ordinal and Nominal Variables 271

12. E. K. Berndt, B. H. Hall, R. E. Hall, and J. A. Hausman. Estimation and

inference in nonlinear structural models. Annals of Economic and Social Mea-

surement, 3:653–665, 1974.

13. C. R. Bhat. Quasi-random maximum simulated likelihood estimation of the

mixed multinomial logit model. Transportation Research, Part B, 35:677–693,

2001.

14. R. D. Bock. Estimating item parameters and latent ability when responses are

scored in two or more nominal categories. Psychometrika, 37:29–51, 1972.

15. R. D. Bock. Multivariate Statistical Methods in Behavioral Research. McGraw-

Hill, New York, 1975.

16. R. D. Bock and M. Aitkin. Marginal maximum likelihood estimation of item

parameters: An application of the EM algorithm. Psychometrika, 46:443–459,

1981.

17. R. D. Bock and M. Lieberman. Fitting a response model for n dichotomously

scored items. Psychometrika, 35:179–197, 1970.

18. R. D. Bock and S. Shilling. High-dimensional full-information item factor

analysis. In M. Berkane, editor, Latent Variable Modeling and Applications to

Causality, pages 163–176. Springer, New York, 1997.

19. N. E. Breslow and X. Lin. Bias correction in generalised linear mixed models

with a single component of dispersion. Biometrika, 82:81–91, 1995.

20. G. Camilli and L. A. Shepard. Methods for Identifying Biased Test Items. Sage,

Thousand Oaks, CA, 1994.

21. M. R. Conaway. Analysis of repeated categorical measurements with condi-

tional likelihood methods. Journal of the American Statistical Association, 84:

53–61, 1989.

22. C. Corcoran, B. Coull, and A. Patel. EGRET for Windows User Manual.

CYTEL Software Corporation, Cambridge, MA, 1999.

23. L. A. Cupples, R. B. D’Agostino, K. Anderson, and W. B. Kannel. Compari-

son of baseline and repeated measure covariate techniques in the Framingham

Heart Study. Statistics in Medicine, 7:205–218, 1985.

24. R. B. D’Agostino, M.-L. Lee, A. J. Belanger, L. A. Cupples, K. Anderson, and

W. B. Kannel. Relation of pooled logistic regression to time dependent Cox

regression analysis: The Framingham Heart Study. Statistics in Medicine, 9:

1501–1515, 1990.

25. M. J. Daniels and C. Gatsonis. Hierarchical polytomous regression models with

applications to health services research. Statistics in Medicine, 16:2311–2325,

1997.

26. K. A. Doksum and M. Gasko. On a correspondence between models in binary

regression analysis and in survival analysis. International Statistical Review,

58:243–252, 1990.

27. D. M. Drukker. Maximum simulated likelihood: Introduction to a special issue.

The Stata Journal, 6:153–155, 2006.

28. J. Engel. On the analysis of grouped extreme-value data with GLIM. Applied

Statistics, 42:633–640, 1993.

29. F. Ezzet and J. Whitehead. A random effects model for ordinal responses from

a crossover trial. Statistics in Medicine, 10:901–907, 1991.

272 Hedeker

30. A. Fielding, M. Yang, and H. Goldstein. Multilevel ordinal models for exami-

nation grades. Statistical Modelling, 3:127–153, 2003.

31. G. M. Fitzmaurice, N. M. Laird, and A. G. Rotnitzky. Regression models for

discrete longitudinal responses. Statistical Science, 8:284–309, 1993.

32. S. A. Freels, R. B. Warnecke, T. P. Johnson, and B. R. Flay. Evaluation of

the effects of a smoking cessation intervention using the multilevel thresholds

of change model. Evaluation Review, 26:40–58, 2002.

33. S. Geman and D. Geman. Stochastic relaxation, Gibbs distributions and the

Bayesian restoration of images. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 6:721–741, 1984.

34. R. D. Gibbons and R. D. Bock. Trend in correlated proportions. Psychome-

trika, 52:113–124, 1987.

35. R. D. Gibbons and D. Hedeker. Random-effects probit and logistic regression

models for three-level data. Biometrics, 53:1527–1537, 1997.

36. G. Glasgow. Mixed logit models for multiparty elections. Political Analysis, 9:

116–136, 2001.

37. H. Goldstein. Nonlinear multilevel models, with an application to discrete

response data. Biometrika, 78:45–51, 1991.

38. H. Goldstein. Multilevel Statistical Models. Edward Arnold, London, 3rd

edition, 2003.

39. H. Goldstein and J. Rasbash. Improved approximations for multilevel models

with binary responses. Journal of the Royal Statistical Society, Series A, 159:

505–513, 1996.

40. W. H. Greene. LIMDEP Version 8.0 User’s Manual. Econometric Software,

Plainview, NY, 4th edition, 2002.

41. P. Haan and A. Uhlendorff. Estimation of multinomial logit models with unob-

served heterogeneity using maximum simulated likelihood. The Stata Journal,

6:229–245, 2006.

42. A. Han and J. A. Hausman. Flexible parametric estimation of duration and

competing risk models. Journal of Applied Econometrics, 5:1–28, 1990.

43. J. Hartzel, A. Agresti, and B. Caffo. Multinomial logit random effects models.

Statistical Modelling, 1:81–102, 2001.

44. D. A. Harville and R. W. Mee. A mixed-model procedure for analyzing ordered

categorical data. Biometrics, 40:393–408, 1984.

45. P. J. Heagerty and S. L. Zeger. Marginalized multilevel models and likelihood

inference. Statisical Science, 15:1–26, 2000. (with discussion)

46. D. Hedeker. MIXNO: A computer program for mixed-effects nominal logistic

regression. Journal of Statistical Software, 4(5):1–92, 1999.

47. D. Hedeker. A mixed-effects multinomial logistic regression model. Statistics

in Medicine, 21:1433–1446, 2003.

48. D. Hedeker and R. D. Gibbons. A random-effects ordinal regression model for

multilevel analysis. Biometrics, 50:933–944, 1994.

49. D. Hedeker and R. D. Gibbons. MIXOR: A computer program for mixed-effects

ordinal regression analysis. Computer Methods and Programs in Biomedicine,

49:157–176, 1996.

6 Multilevel Models for Ordinal and Nominal Variables 273

50. D. Hedeker and R. D. Gibbons. Application of random-effects pattern-mixture

models for missing data in longitudinal studies. Psychological Methods, 2:64–78,

1997.

51.